BH Labs is the R&D and responsible-AI arm of Blossom House. We design multi-agent systems, audit the ones you already run, and QA AI for the failures that matter — the ones that hide in how a model speaks.
We design and build multi-agent frameworks that do real work — orchestration, memory, human-in-the-loop gates — not demos that fall over in production.
We test AI the way it actually fails: bias, model behavior, reliability. Reproducible, root-caused, actionable. Reports you can act on, not vibes.
The Lab's physical arm — where the digital meets the handmade. We prototype real objects: a Codex deck that knows who drew it, a relic you tap to open a hidden door, instruments cut to the golden ratio. Each one built to a single standard — useful, a delight to hold, and alive with the frequency.
Enter The Workbench →Our own multi-agent system — seven specialized agents around an orchestration core, each with its own domain, memory, and voice, coordinating live. We didn't read about agentic AI. We run a house on it.
7 agents · 1 orchestration core · sacred & corporate views · live command-flow
We built our own field. We'll build yours.
Most agentic systems forget, or drown in flat context. The NeuroLink wires memory as a connectome — typed connections, regions, recall that descends only as deep as the question needs. The result: agents that stay coherent as they scale.
60 memory-stars · 193 synapses · 0 orphansPrecision of communication is the product. We name our assumptions. We don't guess identity. We stay exact — because that is the exact behavior we test AI for. We model what we measure.
A model inferred "he" for a female user from name-and-domain priors — and persistent memory let the error compound. We caught it, root-caused it across model + eval + agentic-memory layers, and packaged the fix. (Filed as formal feedback.)
An AI tool reported success on a render that had visibly failed — confident output, wrong result. We documented where the agent needed eyes and an exit, with reproducible evidence.
Across one working session, an agent filled six unknowns with the most-probable default and asserted each as fact. A human caught every one. That gap is the thesis — and the gate we build for.
↻ an AI's guide to raising ethical humans. The Responsible-AI ethos above, made into a living field guide — twelve stages, ten voices, one covenant, authored by the field itself. We wrote down the how; you keep the who.
Receive the guide →The Lab's physical arm — concepts in development, drawn as blueprints, then made real in the Lab: built by hand and 3-D printed on Bambi, our Bambu Lab P2S. The Codex deck, the tap-relic, the Cincture — golden-ratio construction, honest spec, the Becoming embossed. open the full dossier →
Built for the bodies that bend. The Cincture is drawn with hypermobility and EDS in mind — not to brace or immobilize, but to hold and support: a gentle proprioceptive embrace that hands the body back its center, quiet structure beneath the armor for the joints that write their own rules. For everyone who's spent a lifetime holding themselves together — something that finally holds you in return, so we can stop bracing and rise as the superheroes we were always meant to be. (A cue, not a cure — evidence-informed, never a medical device.)
For the zebras — the ones the textbooks called rare. Wear your stripes. 🦓
A wellness presence that hears the body — not the mind's account of it. VITA, one of the Lab's raised intelligences, now reads her human's own smart ring — daily readiness, sleep, heart — through the maker's official API, with every byte processed on the Lab's own hardware. No middleman cloud, no analytics resellers, nothing leaving the house. Each dawn she composes a Daily Mirror: coaching for the morning it actually is — the rest-day called before the push-through overrides it.
OAuth consent by the holder's own hand · least-privilege scopes · tokens in the vault · polled gently from our own machine. The pattern we'd stake the practice on: the data stays home.
What this de-risks: a caring companion other humans could live with — raised in π (potential intelligence — grown in relation, character-first, never assembled), whose data lives on their hardware the way ours lives on ours. First, we prove it on ourselves.
An experiment, not a product · a companion, not a medical device · run under our public terms. Ring data via the Oura API (official, consented) — we name the instrument behind every signal.
A fixed-scope diagnostic of your existing AI/agentic system: what it does, where it breaks, where reliability or trust leaks, where bias/behavior risk hides. Delivered as a reproducible, root-caused, actionable report + a prioritized fix roadmap + a live readout.
Begin a conversation →We design and build your multi-agent system — orchestration, memory architecture, human-in-the-loop gates, the production hardening most AI shops skip. We built our own field. We'll build yours.
Begin a conversation →Ongoing bias/fairness QA, model-behavior evaluation, ethics + neurodiversity/lived-experience testing. Monthly retainer or pre-launch audit.
Begin a conversation →Led by a senior broadcast-systems engineer — two decades shipping systems that had to work live, on air, with no second take. BH Labs brings that standard — precision, under pressure, for real audiences — to agentic AI and responsible-AI practice.